**3. Probabilistic model for time-series analysis**

Understanding the pattern of complications associated with the disease has been used significantly in the clinical domain [6]. It provides an insight into the prediction and relative prevention of the associated complications which are expected to occur in a patient follow-up [7]. It generally can lead to less suffering time for patients while saving time and cost to healthcare. However, that is highly dependent on the stage of disease along with the prior occurring complications, which is associated with time-series analysis. In time-series analysis, every disease risk factor and complication is determined by various features in previous patient visits (time interval). At every medical visit, all diabetic patients have a unique profile of symptoms and complications that change over time, regardless of the phase of the disease. This non-stationary characteristic of clinical data collected as part of the monitoring of T2DM creates a difficult context for effective forecasting [8]. Clinical data needs to be considered as time-series data in order to provide a description of the progression of a disease over time. Nevertheless, dealing with time-series patient records is known to be a major issue in the prognosis of comorbidities [9], particularly when time-series data is imbalanced and contains few examples of patients without comorbidities that are common to all patients. In Type 2 Diabetes, for example, once patients are diagnosed with T2DM, half of them show signs of complications [10]. Unfortunately, these life-threatening complications remain undiagnosed for a long time because of the hidden patterns of their associated risk factors [11]. If T2DM is not appropriately managed, the development of serious complications, such as neuropathy, retinopathy, and hypertension lead to disability, premature mortality and financial cost [12]. The prediction process is complex due to the interactions between these complications and other features, as well as between complications themselves. More importantly, each patient has a unique profile of complication occurrence and the status of T2DM risk factors during a patients time-series is subject to change, as their levels may rise and fall over time. Early diagnosis and prevention techniques are needed to reduce the associated mortality and morbidity caused by T2DM complications [13]. Although there are various methodologies for T2DM prediction, for example, risk-prediction equation and Markov models [14], studies that enable early predictions of diabetes using predictive models are limited [15]. The risk-prediction equations suffer from uncertainty as well as performing only one-step-ahead predictions, while Markov models are limited to a small number of discrete risk factors. Other existing literature on investigating the prognosis of T2DM complications [16, 17] focuses particularly on logistic regression and Naive Bayes. Such studies are unsatisfactory for modelling the complex T2DM complications/risk factors. Logistic regression does

**1. Introduction**

*Type 2 Diabetes - From Pathophysiology to Cyber Systems*

Intelligent systems, whether biological or artificial, require the ability to make decisions under uncertainty using the available evidence. Several computational models exhibit some of the required functionality to handle uncertainty. These computational models in Artificial Intelligent (AI) and Machine Learning are judged by two main criteria: ease of creation and effectiveness in decision making. For example, Neural Networks (NNs) which represent complex input/output relations using combinations of simple nonlinear processing elements, are a familiar tool in AI and computational neuroscience. Alternatively, probabilistic networks (also called Bayesian Networks) are a more explicit representation of a domain through modelling the joint probability distribution (the probability of all possible outcomes in a domain). This paper provides a short summary of the previous methods in Intelligent Data Analysis (IDA) in disease progression, decision making and probabilistic modelling of patients. It then describes some existing key methods that can be updated or combined to model multiple diabetes complications in the presence of unmeasured factors. There is considerable research on predicting diabetes, especially Type 2 Diabetes Mellitus (T2DM), complications. Nevertheless, the previous research of the author is discussed in order to address these issues. In particular, these previously proposed methods have contributed to the diabetes literature by explaining unknown risk factors and identifying temporal phenotypes employing hybrid methods (including descriptive and predictive). These suggested methodology includes rule-based methods for an explanation of patient subgroups

and a probabilistic framework for modelling data explicitly.

**progression modelling**

progression [5].

**198**

**2. Literature review: intelligent data analysis in complex disease**

This article reviews the current literature on some of the most common AI methodologies, including probabilistic modelling, association rule mining, and latent variable discovery. Intelligent Data Analysis (IDA) is a subcategory of AI that is focused on data analysis and modelling. These methods are known to be highly successful in combining advantages of modern data analytics, classical statistics and the expertise of scientists and experts [1–3]. IDA techniques have already proved successful in clinical modelling [4]. A large and growing body of literature has investigated IDA approaches that have shown excellent results modelling crosssectional clinical data for classification. There has also been substantial modelling on longitudinal data using IDA techniques. However, there is still an urgent need to improve these models to take account of the variability of disease progression from person to person, and explicitly model the time-varying nature of the disease. Many studies have attempted to find automated ways of helping clinicians predict disease

For many clinical problems, the underlying structure of unmeasured variables

unmeasured information such as social deprivation, missing clinical data, environmental factors, time-based information or some combination of these. To gain trust in any AI model, it is mandatory to understand/explain influencing factors of

may play an essential role in the progress of the disease. However, it is still a relatively unexplored area. Identifying these unmeasured variables as hidden or latent variables is key. What is more, understanding the semantics behind these unmeasured risk factors can improve the understanding of the disease mechanisms and thus better improve clinical decision making. Interpreting these latent variables

is complicated; however, as they may represent different many types of

risk factors. Many diseases involved structural changes based upon key stages in the progression, but many models did not appear to take this into account. There has been some work in extending DBNs to model underlying processes that are nonstationary [24]. In [24], clinical features were modelled using a second-order timeseries model while time-invariant temporal dependencies were assumed. Among this, some studies, for example, Marini and co-authors conducted research [23] that variables were connected within two-time-series and within the same time slice assumed that the temporal dependencies were time-invariant. In addition, in Marinis paper for learning the network structures, a Tabu search was used based on the Hill climbing algorithm for Bayesian Networks but with no use of latent variables. However, the approach was useful for stratifying patients according to the probability of developing complications, the major limitation of the Marinis work

*Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial…*

Another work in [25] retained the stationary nature of the structure in favour of parameter flexibility, arguing that structure changes lead almost certainly to overflexibility of the model in short time-series. Alternatively, a paper [26] formalised non-stationary DBN models and suggested MCMC sampling algorithm for learning the structure of the model from time-series biological data. Similarly, authors in [27] estimated the variance in the data structure parameter with an MCMC approach, but the search space was limited to a fixed number of segments and indirect edges only, which is not suitable for T2DM data. Such studies remained narrow and limited by constraints on one or more degrees of freedom: the segmentation points of the time-series, the parameters of the variables, the dependencies between the variables and the number of segments and the ignorance of the incom-

Another common problem with classifying complications in longitudinal data is

that there may be many more cases where the complication does not manifest compared to those where it does. Early prediction of T2DM complications while discovering the behaviour of associated aggressive risk factors can help to improve a patients quality of life [28]. This study suggests that while there is an association between the latent variable and joint complications in the prognosis of T2DM patients, this relationship is complex. In T2DM data analysis, another challenge can be to classify/group patients in imbalanced clinical data with several binary complications. Models of the time-series data are needed to manage diabetic complications and deal with their imbalanced and complex interactions. In particular, mining time-series is one of the challenging problems in the prognosis of disease. In addition, it has received considerable critical attention in data mining especially when there are rare positive results [29]. It has been reported that a class imbalance in the training data caused by one class (here positive cases) massively outnumbers the examples in another class (negative class) [30]. This situation may occur where the number of positive clinical test results for a complication is not equal or even close to the number of negatives. That can be solved by applying an appropriate balancing strategy in a multi-class classification problem. Different learning techniques deal with imbalanced data, such as oversampling, undersampling, boosting, bagging, bootstrapping, and repeated random sub-sampling [31]. Therefore, this chapter in order to prepare T2DM data for the prediction has utilised these strategies and customised them based on dataset nature (time-series patients records with the unequal number of visits). As a result, various balancing strategies such as pair-sampling, bootstrapping undersampling and over-sampling have

derived from time discretisation in time slices of one year.

*DOI: http://dx.doi.org/10.5772/intechopen.94228*

plete data and latent variable.

been proposed in [32, 33].

**201**

**3.2 Dealing with time-series imbalanced data**

#### **Figure 1.**

*The organs/muscles affected by the common complications associated with type 2 Diabetes.*

not perform well when there are multiple or non-linear decision limitations. In Naive Bayes, there is an assumption of independence among the risk factors whereas all features are independent of one another.

The major limitation of the previous work in T2DM literature derives from time discretisation in temporal time slices per year. Therefore, in this study, we consider all T2DM patient's follow-up visits regardless of year basis while precisely monitoring the location of change within the unequal number of visits. This chapter suggests that AI in Medicine can provide useful techniques to analyse patient data to be able to find cure for the disease or reduce patient's suffering time (see **Figure 1**).

#### **3.1 Dynamic Bayesian networks**

In the field of medical informatics, probabilistic IDA techniques are exploited to obtain different clinical solutions. To improve patients' quality of life, there is an urgent need to extend and explore probabilistic IDA methods to investigate the disease complications from a clinical point of view. Thus, a Bayesian Network (BN) decision model was exploited in [18] for supporting the diagnosis of dementia, Alzheimer disease and mild cognitive impairment. Bayesian Network models appear to be well suited T2DM progression modelling, because of their flexibility in modelling spatial and temporal relationships as well as their ease of interpretation [19]. It has been reported that Dynamic Bayesian Networks (DBNs) are simple BNs for modelling time-series data and popular for modelling uncertain noisy timeseries clinical data [20]. More importantly, DBNs are probabilistic graphical models that can handle missing data and hidden variables.

Previous work on learning DBNs have inferred both network structures and parameters from (sometimes incomplete) clinical datasets [20]. For example, a recent study presented a DBN method but to analyse fisheries data [21]. Authors in [22] proposed a Bayes Network to predict diabetes on the Pima Indian Diabetes dataset. However, the study failed to consider the time-series analysis. Similarly, authors in another study [23] simulated the health state and complications of type 1 diabetes patients by using partially and entirely learned Bayesian models. Apart from using a different type of Diabetes, this chapter is utilising a different approach from the above studies for the representation of the relationship between T2DM

#### *Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial… DOI: http://dx.doi.org/10.5772/intechopen.94228*

risk factors. Many diseases involved structural changes based upon key stages in the progression, but many models did not appear to take this into account. There has been some work in extending DBNs to model underlying processes that are nonstationary [24]. In [24], clinical features were modelled using a second-order timeseries model while time-invariant temporal dependencies were assumed. Among this, some studies, for example, Marini and co-authors conducted research [23] that variables were connected within two-time-series and within the same time slice assumed that the temporal dependencies were time-invariant. In addition, in Marinis paper for learning the network structures, a Tabu search was used based on the Hill climbing algorithm for Bayesian Networks but with no use of latent variables. However, the approach was useful for stratifying patients according to the probability of developing complications, the major limitation of the Marinis work derived from time discretisation in time slices of one year.

Another work in [25] retained the stationary nature of the structure in favour of parameter flexibility, arguing that structure changes lead almost certainly to overflexibility of the model in short time-series. Alternatively, a paper [26] formalised non-stationary DBN models and suggested MCMC sampling algorithm for learning the structure of the model from time-series biological data. Similarly, authors in [27] estimated the variance in the data structure parameter with an MCMC approach, but the search space was limited to a fixed number of segments and indirect edges only, which is not suitable for T2DM data. Such studies remained narrow and limited by constraints on one or more degrees of freedom: the segmentation points of the time-series, the parameters of the variables, the dependencies between the variables and the number of segments and the ignorance of the incomplete data and latent variable.

## **3.2 Dealing with time-series imbalanced data**

Another common problem with classifying complications in longitudinal data is that there may be many more cases where the complication does not manifest compared to those where it does. Early prediction of T2DM complications while discovering the behaviour of associated aggressive risk factors can help to improve a patients quality of life [28]. This study suggests that while there is an association between the latent variable and joint complications in the prognosis of T2DM patients, this relationship is complex. In T2DM data analysis, another challenge can be to classify/group patients in imbalanced clinical data with several binary complications. Models of the time-series data are needed to manage diabetic complications and deal with their imbalanced and complex interactions. In particular, mining time-series is one of the challenging problems in the prognosis of disease. In addition, it has received considerable critical attention in data mining especially when there are rare positive results [29]. It has been reported that a class imbalance in the training data caused by one class (here positive cases) massively outnumbers the examples in another class (negative class) [30]. This situation may occur where the number of positive clinical test results for a complication is not equal or even close to the number of negatives. That can be solved by applying an appropriate balancing strategy in a multi-class classification problem. Different learning techniques deal with imbalanced data, such as oversampling, undersampling, boosting, bagging, bootstrapping, and repeated random sub-sampling [31]. Therefore, this chapter in order to prepare T2DM data for the prediction has utilised these strategies and customised them based on dataset nature (time-series patients records with the unequal number of visits). As a result, various balancing strategies such as pair-sampling, bootstrapping undersampling and over-sampling have been proposed in [32, 33].

not perform well when there are multiple or non-linear decision limitations. In Naive Bayes, there is an assumption of independence among the risk factors

*The organs/muscles affected by the common complications associated with type 2 Diabetes.*

The major limitation of the previous work in T2DM literature derives from time discretisation in temporal time slices per year. Therefore, in this study, we consider all T2DM patient's follow-up visits regardless of year basis while precisely monitoring the location of change within the unequal number of visits. This chapter suggests that AI in Medicine can provide useful techniques to analyse patient data to be able to find cure for the disease or reduce patient's suffering time (see **Figure 1**).

In the field of medical informatics, probabilistic IDA techniques are exploited to obtain different clinical solutions. To improve patients' quality of life, there is an urgent need to extend and explore probabilistic IDA methods to investigate the disease complications from a clinical point of view. Thus, a Bayesian Network (BN) decision model was exploited in [18] for supporting the diagnosis of dementia, Alzheimer disease and mild cognitive impairment. Bayesian Network models appear to be well suited T2DM progression modelling, because of their flexibility in modelling spatial and temporal relationships as well as their ease of interpretation [19]. It has been reported that Dynamic Bayesian Networks (DBNs) are simple BNs for modelling time-series data and popular for modelling uncertain noisy timeseries clinical data [20]. More importantly, DBNs are probabilistic graphical models

Previous work on learning DBNs have inferred both network structures and parameters from (sometimes incomplete) clinical datasets [20]. For example, a recent study presented a DBN method but to analyse fisheries data [21]. Authors in [22] proposed a Bayes Network to predict diabetes on the Pima Indian Diabetes dataset. However, the study failed to consider the time-series analysis. Similarly, authors in another study [23] simulated the health state and complications of type 1 diabetes patients by using partially and entirely learned Bayesian models. Apart from using a different type of Diabetes, this chapter is utilising a different approach from the above studies for the representation of the relationship between T2DM

whereas all features are independent of one another.

*Type 2 Diabetes - From Pathophysiology to Cyber Systems*

that can handle missing data and hidden variables.

**3.1 Dynamic Bayesian networks**

**Figure 1.**

**200**

The bootstrap approach can be used to identify the significant statistics from classifiers learnt from such data. For example, in a study [34], Li and co-authors provide an extension to the temporal bootstrap approach while applied on crosssectional data. Similarly, a study conducted in [21], the bootstrap strategy is extended to longitudinal data by sampling pairs of time points, thus enabling the (first-order) temporal nature of the data to be inferred. However, these solutions only can be applicable when the imbalance ratio for all binary complications is similar. Otherwise, it can be more difficult if we need to over-sample one class value and under- sample others in order to reduce bias from data. Overall, the observed balancing strategies from the prior studies have not been sufficient for analysing more than one complication at a time, whereas it was almost impossible to obtain a satisfactory prediction performance enhancement for all complications. As well as modelling unmeasured factors, hidden variables can also be used to model nonstationary processes. This chapter attempts to address this issue by using hidden variables discovery approaches based upon T2DM risk factors/complications dependencies. Before explaining these strategies, it is necessary to understand unmeasured variables and analyse their dependencies that are generated by causal structures.

consequence there has been some controversy about that status of exploratory versus confirmatory factor analysis. In this regard, casual discovery methods in AI have the advantages as they can discover the actual dependencies and indepen-

*Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial…*

The causal discovery of BNs is a critical research territory, which depends on looking through the space of causal models for those which can best clarify a pattern of probabilistic conditions appeared in the data [35]. As a result, [38] showed the integration of structure-search algorithm with a latent variable in a DBNs model. However, the method did not consider the discovery of the long-range dependencies with an equal number of time slices. Similarly, in [39], Bayesian belief networks was used to find the most probable structure, using the K2 algorithm, while adding a hidden variable. Nevertheless, Cooper in [39] applied the K2 method that needs an ordering on the nodes. Witting focused on using hidden variables in a known structure [40]. Cooper in [39] used Bayesian techniques to find the most probable structure and can use this technique to add hidden variables. In principle, exact Bayesian methods for hidden variables could identify the most probable structures of factors given the data and suitable priors. However, with a large number of variables, exact methods are prohibitively expensive. Furthermore, in [41] Silva highlighted the weakness of DAG (Directed Acyclic Graph) models in the

marginalisation of Hidden factors and representing the independencies over a subset of features in a DAG with more links. They suggested that Directed mixed graphs (DMGs) are a solution to this drawback. Therefore, they represented how to perform Bayesian inference on two DMGs, such as Gaussian and Probit, which is not the focus of this chapter. Nevertheless, such studies remained narrow and limited by constraints on one or more degrees of freedom: the segmentation points of the time-series, the parameters of the variables, the dependencies between the variables and the number of hidden factors. As a result, Chicharro in [42] analysed causal influences to find the relationship among different brain regions in several disorders. Similar to this chapter, Chicharros research made use of Inductive Causation (IC\*) algorithm in the latent process to analyse Granger causality and Dynamic Causal Modelling. However, Chicharros study did not consider DBNs to

Difficulties arise, however, when an attempt is made to implement a Bayesian Network structure as authors in [43] have argued that the number of potential DAGs over the disease risk factors is super-exponential. Additionally, the real cause-effect relationship DAG is not distinguishable while from equivalent structures when learning only using from observational data. This issue will be worse, especially when each expert has a unique probability of correctly labelling the inclusion or exclusion of edges in the disease structure. As noted by Amirkhani [43], some scoring functions are provided with that score each suitable graph based on the data and experts knowledge. Another research in [44] shows that networks with the fixed structure containing hidden variables can be learned automatically from data using a gradient-descent mechanism similar to that used in neural networks. A few algorithms have been created to understand the structure for Bayesian Networks from both fully observed models and those with hidden variables. Structure Expectation–Maximisation (SEM) has been produced for learning Probabilistic system structure from information with latent factors and missing data. A structure learning algorithm has been created for non-stationary dynamic probabilistic models. For example, REVEAL (REVerse Engineering ALgorithm) has been utilised as a structure learning algorithm, that learns the optimal set of parents for each node of a network independently, based on the information-theoretic concepts of mutual information analysis. However, the two-stage temporal Bayes network (2TBN) cannot be well recovered by the application of REVEAL. A normally

dencies in the data.

*DOI: http://dx.doi.org/10.5772/intechopen.94228*

understand causal influences.

**203**

#### **3.3 Causal structure learning and latent variable discovery**

Various studies on longitudinal data sets have suggested an association between complications and risk factors of the disease. To discover probabilistic dependencies given clinical data, it is necessary to search the space of belief networks or casual models, which is called casual discovery of BNs [35]. These patterns of dependency with no model based solely upon the observed variables can be explained by using a latent variable. The casual discovery indicates dependencies that are generated by casual structures with unmeasured factors, i.e., hidden variables. Hidden variable modelling, introduced in [36], has a long tradition in casual discovery. One of the research gap in the previous literature of disease prediction is the existence of the unmeasured or latent variables. This is because clinicians cannot measure all risk factors and carry out all kinds of tests, so there are some unmeasured factors that clinicians fail to measure, which need to be discovered at the early stage of diabetes.

Furthermore, Factor Learning (FL) was introduced in [37], which has been known as one method for learning a probabilistic model from data. It can also be helpful to understand latent variables and measure their hypothetical impacts. FL contrasts with most other BN learning methods in that it learns a factor structure. As Martin and co-authors in [37] stated that FL for hidden variables could identify the most probable structures of factors have given the data and suitable priors. However, with a large number of variables, FL methods might be prohibitively expensive. Again in the same research these authors provided a factor structure for learning methods that efficiently utilised hidden variables. Factor structure indicates the joint probability distribution among discrete observed variables. It also contributes an explanation across a small number of variables. Although factor structures are suitable for polynomial time inference, they can cause a reduction in the prediction accuracy and precision; they contribute an explanation across a small number of variables. Nevertheless, these techniques failed to consider prior belief in the factor structure, and therefore, it could be hard to rely on the final structure.

Factor structure indicates the joint probability distribution among discrete observed variables. Interestingly, each factor in a factor structure corresponds to a completely connected dependency graph. Although they are suitable for polynomial time inference, caused reducing accuracy and precision. By contrast, they are not able to decide precisely whether or not latent variables are present, and in

*Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial… DOI: http://dx.doi.org/10.5772/intechopen.94228*

consequence there has been some controversy about that status of exploratory versus confirmatory factor analysis. In this regard, casual discovery methods in AI have the advantages as they can discover the actual dependencies and independencies in the data.

The causal discovery of BNs is a critical research territory, which depends on looking through the space of causal models for those which can best clarify a pattern of probabilistic conditions appeared in the data [35]. As a result, [38] showed the integration of structure-search algorithm with a latent variable in a DBNs model. However, the method did not consider the discovery of the long-range dependencies with an equal number of time slices. Similarly, in [39], Bayesian belief networks was used to find the most probable structure, using the K2 algorithm, while adding a hidden variable. Nevertheless, Cooper in [39] applied the K2 method that needs an ordering on the nodes. Witting focused on using hidden variables in a known structure [40]. Cooper in [39] used Bayesian techniques to find the most probable structure and can use this technique to add hidden variables. In principle, exact Bayesian methods for hidden variables could identify the most probable structures of factors given the data and suitable priors. However, with a large number of variables, exact methods are prohibitively expensive. Furthermore, in [41] Silva highlighted the weakness of DAG (Directed Acyclic Graph) models in the marginalisation of Hidden factors and representing the independencies over a subset of features in a DAG with more links. They suggested that Directed mixed graphs (DMGs) are a solution to this drawback. Therefore, they represented how to perform Bayesian inference on two DMGs, such as Gaussian and Probit, which is not the focus of this chapter. Nevertheless, such studies remained narrow and limited by constraints on one or more degrees of freedom: the segmentation points of the time-series, the parameters of the variables, the dependencies between the variables and the number of hidden factors. As a result, Chicharro in [42] analysed causal influences to find the relationship among different brain regions in several disorders. Similar to this chapter, Chicharros research made use of Inductive Causation (IC\*) algorithm in the latent process to analyse Granger causality and Dynamic Causal Modelling. However, Chicharros study did not consider DBNs to understand causal influences.

Difficulties arise, however, when an attempt is made to implement a Bayesian Network structure as authors in [43] have argued that the number of potential DAGs over the disease risk factors is super-exponential. Additionally, the real cause-effect relationship DAG is not distinguishable while from equivalent structures when learning only using from observational data. This issue will be worse, especially when each expert has a unique probability of correctly labelling the inclusion or exclusion of edges in the disease structure. As noted by Amirkhani [43], some scoring functions are provided with that score each suitable graph based on the data and experts knowledge. Another research in [44] shows that networks with the fixed structure containing hidden variables can be learned automatically from data using a gradient-descent mechanism similar to that used in neural networks. A few algorithms have been created to understand the structure for Bayesian Networks from both fully observed models and those with hidden variables. Structure Expectation–Maximisation (SEM) has been produced for learning Probabilistic system structure from information with latent factors and missing data. A structure learning algorithm has been created for non-stationary dynamic probabilistic models. For example, REVEAL (REVerse Engineering ALgorithm) has been utilised as a structure learning algorithm, that learns the optimal set of parents for each node of a network independently, based on the information-theoretic concepts of mutual information analysis. However, the two-stage temporal Bayes network (2TBN) cannot be well recovered by the application of REVEAL. A normally

The bootstrap approach can be used to identify the significant statistics from classifiers learnt from such data. For example, in a study [34], Li and co-authors provide an extension to the temporal bootstrap approach while applied on crosssectional data. Similarly, a study conducted in [21], the bootstrap strategy is extended to longitudinal data by sampling pairs of time points, thus enabling the (first-order) temporal nature of the data to be inferred. However, these solutions only can be applicable when the imbalance ratio for all binary complications is similar. Otherwise, it can be more difficult if we need to over-sample one class value and under- sample others in order to reduce bias from data. Overall, the observed balancing strategies from the prior studies have not been sufficient for analysing more than one complication at a time, whereas it was almost impossible to obtain a satisfactory prediction performance enhancement for all complications. As well as modelling unmeasured factors, hidden variables can also be used to model nonstationary processes. This chapter attempts to address this issue by using hidden variables discovery approaches based upon T2DM risk factors/complications dependencies. Before explaining these strategies, it is necessary to understand unmeasured variables and analyse their dependencies that are generated by causal

Various studies on longitudinal data sets have suggested an association between complications and risk factors of the disease. To discover probabilistic dependencies given clinical data, it is necessary to search the space of belief networks or casual models, which is called casual discovery of BNs [35]. These patterns of dependency with no model based solely upon the observed variables can be explained by using a latent variable. The casual discovery indicates dependencies that are generated by casual structures with unmeasured factors, i.e., hidden variables. Hidden variable modelling, introduced in [36], has a long tradition in casual discovery. One of the research gap in the previous literature of disease prediction is the existence of the unmeasured or latent variables. This is because clinicians cannot measure all risk factors and carry out all kinds of tests, so there are some unmeasured factors that clinicians fail to measure, which need to be discovered at the early stage of diabetes. Furthermore, Factor Learning (FL) was introduced in [37], which has been known as one method for learning a probabilistic model from data. It can also be helpful to understand latent variables and measure their hypothetical impacts. FL contrasts with most other BN learning methods in that it learns a factor structure. As Martin and co-authors in [37] stated that FL for hidden variables could identify the most probable structures of factors have given the data and suitable priors. However, with a large number of variables, FL methods might be prohibitively expensive. Again in the same research these authors provided a factor structure for learning methods that efficiently utilised hidden variables. Factor structure indicates the joint probability distribution among discrete observed variables. It also contributes an explanation across a small number of variables. Although factor structures are suitable for polynomial time inference, they can cause a reduction in the prediction accuracy and precision; they contribute an explanation across a small number of variables. Nevertheless, these techniques failed to consider prior belief in the factor structure, and therefore, it could be hard to rely on the final structure. Factor structure indicates the joint probability distribution among discrete observed variables. Interestingly, each factor in a factor structure corresponds to a completely connected dependency graph. Although they are suitable for polynomial time inference, caused reducing accuracy and precision. By contrast, they are not able to decide precisely whether or not latent variables are present, and in

**3.3 Causal structure learning and latent variable discovery**

*Type 2 Diabetes - From Pathophysiology to Cyber Systems*

structures.

**202**

utilised structure learning algorithm depends on REVEAL which takes in the ideal arrangement of guardians for every hub of a system autonomously, in light of the theoretical data ideas of common data examination. Be that as it may, the twoarrange fleeting Bayes organise as the 2TBN which cannot be all around recuperated by use of REVEAL. Rijmen in [45] exploited an HMM to study the temporal pattern of symptoms burden in brain tumour patients. He showed that the discovery of symptom experience over time is necessary for treatment and follow-up of patients with symptom-specific intervention. In general, Bayesian learning methods could determine network structure and how the networks variables should be represented along with the causal links among them. Moreover, it addressed the difficulty of qualifying causal relationships in terms of Conditional Probability Tables (CPTs). Witting focused on using hidden variables in a known structure [40] as the knowledge of the latent variable in predictive modelling is important for an understanding of the complex AI models. Discovering latent variables can potentially capture unmeasured effects from clinical data, simplifying complex networks of interactions and giving us a better understanding of disease processes. In addition, it can improve classification accuracy and boost user confidence in the classification models [46]. Elidan and co-authors in [47] emphasised the importance of the presence of hidden variables. In addition, they determined a hidden variable that interacted with observed variables and located them within the Bayesian Network structure. They also showed that networks without hidden variables are clearly less useful because of the increased number of edges needed to model all interactions, which caused overfitting. Despite the productivity of exploring trees of hidden variables to render all observable variables independently [48], these hidden variables were non-optimal with independencies among observable variables. Overall, previous works on learning DBNs have presented both network structures and parameters from clinical data sets and learning parameters for a fixed network of incomplete data, in the presence of missing data and latent variables [20]. Much of the current literature on disease prediction have argued that a complex AI model, with many unexplainable hidden variables, also has several serious drawbacks. Therefore, this chapter has chosen AI DBNs model to learn parameters and latent variables to predict complications. The next section intends to emphasise the explainability of the proposed methodology in order to uncover the meaning behind the latent AI model.

obstacles that arise to interpret the findings, such as the scale of big data, complex

*Predicting Type 2 Diabetes Complications and Personalising Patient Using Artificial…*

Most medical algorithms proposed by [49], such as AI Doctor designed to reproduce current problem-solving methods (e.g., the detection of cancers). In addition, the concept assignment can help people to strengthen their skills and talents for a computer system that showcased superhuman effectiveness and

Google's AI Doctor can be demonstrated how they could be used to provide an explanation further into predictions generated by local classifiers, first from conventional image classification networks to a focused clinical application. The concept attribution approach in AI Doctor offers several promising avenues for future work. In addition to this, the concept assignment can help people to strengthen their skills and talents for a computer system that showcases superhuman effectiveness and efficiency. The concepts of explanatory power are outlined by Google under three principle assumptions/limitations: firstly, comprehension for whatever hidden layer and artificial neurons would offer. This is based on most of the information in a deep neural network consists of hidden layers. Secondly, it recommends that acknowledging the numerous hidden layers and understanding their design on a meta-level would lead to more in-depth modelling. Finally, to comprise how nodes become active, it considers groups of interconnected neurons that trigger at the same time and space. These principles are defined instead of

explaining the structural nature of each neuron in each network. This is because the stratification of a network for the categories of interconnected neurons would enable its configurations even more abstractable. This is the main weakness of the

explainability issues while enabling human-like description of the internal state of a deep network by employing Concept Activation Vectors (CAVs). While medical systems are mostly designed to reproduce current decision-making methods such as the classifier used in the detection of cancers, Google has claimed that its novel strategy can interpret existing clinical data. Although Google has made a claim that the CAVs can directly relate to one's anticipated theories, to draw conclusions about the decision-making process, it needs to consider the human needs of a higher level

Nevertheless, cardiac specialists have been critical of the conclusions derived by Google in the clinical domain. With the proper information, AI is optimistic that innovative, unique healthcare insights might be created without human intervention. Unfortunately, this new approach is only established based on extensive and adequate datasets. This is presumably part of the explanation of why Google has established projects as its benchmark research proposal is capturing detailed patients' history of 100,000 population across four years. However, the investigation conducted out by Google did not necessarily indicate that the suggestion was entirely distant. Such as image classifiers that could be applied to low-level structures. The central concept and assumption are to consider a neural network as additional assistance that can cause issues related to the internal representation. As a

One of the most highlighted ones is Google's approach to resolve the

result, the clinicians commented on the deep explanatory networks. They questioned the hypotheses, by stating that although the AI algorithms and Deep Learning could improve current prediction methods of clinical domain, the research would not be trustworthy unless it had been assessed with caution while a broader range of disease had been explored. Difficulties arose, when an attempt was made in

interactions, and high-dimensional internal state.

*DOI: http://dx.doi.org/10.5772/intechopen.94228*

*4.1.1 Google's novel approach*

efficiency.

black box models.

of understandability.

**205**
